MyKSC: Disaggregated Containerized Supercomputer Platform.

Recently, with the emergence of cloud-based machine learning application services such as ChatGPT and big data analytics, which require high-volume data learning, demand for large-scale high-performance computing (HPC) resources has significantly increased. However, there are many difficulties in providing cloud-based services owing to the differences in existing HPC resource operation and execution environments. To address these challenges, there are many ongoing research efforts to support cloud application services on HPC systems using container technology. However, there are fundamental differences between traditional HPC and cloud applications that make administrators hesitant to adopt them in actual operating environments. To solve these problems, this paper introduces MyKSC, which provides a unified user interface based on a loosely-coupled architecture. MyKSC is a system applied to Nurion, a supercomputer operated by the Korean National Supercomputing Center (KISTI). Nurion users can use new cloud application services such as Jupyter and RStudio along with existing HPC application services like MPI through MyKSC. To achieve this, MyKSC divides the entire system into Kubernetes and HPC clusters and selects suitable cluster resources to run the application based on user preferences. This paper proposes the loosely-coupled containerized supercomputer platform and introduces its current implementation.